An Online Framework for Learning Novel Concepts over Multiple Cues

Abstract

Abstract. We propose an online learning algorithm to tackle the problem of learning under limited computational resources in a teacher-student scenario, over multiple visual cues. For each separate cue, we train an online learning algorithm that sacrifices performance in favor of bounded memory growth and fast update of the solution. We then recover back performance by using multiple cues in the online setting. To this end, we use a two-layers structure. In the first layer, we use a budget online learning algorithm for each single cue. Thus, each classifier provides confidence interpretations for target categories. On top of these classifiers, a linear online learning algorithm is added to learn the combination of these cues. As in standard online learning setups, the learning takes place in rounds. On each round, a new hypothesis is estimated as a function of the previous one. We test our algorithm on two student-teacher experimental scenarios and in both cases results show that the algorithm learns the new concepts in real time and generalizes well.

OM-2: An online multi-class Multi-Kernel Learning algorithm Luo Jie

“…We compared the performance of our algorithm to the OMCL algorithm [15], to the Passive-Aggressive algorithm (PA-I) [7] using the average kernel and the single best feature, and to a batch MKL algorithm [25] 2 . We determined all of our online learning parameters via cross-validation.…”

“…We determined all of our online learning parameters via cross-validation. The features and kernel parameters were obtained directly from the authors who made this information publicly available [15,11].…”

“…This dataset is ideal for testing online learning algorithm: the algorithm has to incrementally update the model, so to adapt to the variations captured in the dataset. For experiments, we used the same setup described in [16,15]. We considered the 12 sequences acquired by robot Dumbo, and divided them into training and test sets, where each training sequence has a corresponding one in the test sets captured under roughly similar conditions.…”

“…In total, we considered twelve different permutations of training and test sets. The images were described using three visual descriptors and a geometric feature from the Laser Scan sensor, as in [15]. Figure 2 reports the average online training and recognition error rate on the test set.…”

“…al. [15] present a wrapper algorithm using a two-layer structure, which can use most of the known online learning methods as base algorithms. The specific version proposed in [15] has a bounded memory complexity, but there is neither mistake bound nor convergence guarantee.…”